mlx-community/Qwen3-TTS-12Hz-1.7B-CustomVoice-4bit

Text to speech with vllm-omni-mlx on Apple M4 (16 GB unified memory).

Text to speech4-bitvllm-omni-mlx1 × M4 (16 GB unified memory)unverified

INSTALLvllm-omni-mlx · MLXShell

Run in the prepared environment

git clone https://github.com/InferMatrixLab/vllm-omni-mlx.git
cd vllm-omni-mlx
git checkout 210cc4e35e208b56d4146e7cc735952cde8ea24e
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[tts]'

vllm-omni-mlx serve mlx-community/Qwen3-TTS-12Hz-1.7B-CustomVoice-4bit --omni --api-key demo

HARDWARE

NVIDIA

Huawei Ascend

Apple Silicon

VARIANT

4bit

mlx-community/Qwen3-TTS-12Hz-1.7B-CustomVoice-4bit

ENGINE

BACKEND

macOS arm64

TASK

Single device · recorded configuration

VERIFICATION

Deployment unverifiedAcceptance INCONCLUSIVE

Not revalidated for this catalog. See the original evidence for the exact scope.

ADVANCED Source & reproducibility

Deployment ref: 210cc4e35e208b56d4146e7cc735952cde8ea24e
Recipe source: InferMatrixLab/vllm-omni-mlx/recipes/qwen3-tts-mlx-m4-4bit.yaml
Content commit: 4bf0a5ccea1d5bf364f62927bf015368879a68a1
Updated: 2026-10-19
Maintainers: @hsliuustc0106

Edit the source recipe

Environment & prerequisites

macOS arm64

One Apple M4 with 16 GB unified memory, as documented in the Qwen3-TTS spike; exact macOS release is not recorded.
Python 3.10+; the recorded spike used Python 3.13.5, mlx 0.32.3 and mlx-audio 0.5.7.
Install the source [tts] extra; MLX requires Apple Silicon arm64 wheels.
Initial source install and model load require network access and disk space for the model snapshot.

Limitations & verification scope

This recipe has not been re-run on Apple hardware during the 2026-10-19 catalog work.
M4/16 GB and the exact checkpoint are documented in the historical spike, which measured an earlier mlx-audio adaptation rather than this deployment ref.
Source for the hardware scope: https://github.com/InferMatrixLab/vllm-omni-mlx/blob/210cc4e35e208b56d4146e7cc735952cde8ea24e/docs/spike-mlx-audio-qwen3-tts.md
The dependency ranges and model download are not a fully locked macOS environment or checkpoint revision. Record resolved versions, model snapshot and macOS release during remote validation.
The 4-bit label describes the quantized checkpoint; embeddings and vocoder may retain higher precision.
One model instance per process; generation is serialized. The demo API key is a local example; default host is 127.0.0.1.
No claim for every Apple Silicon SKU, ASR, speech-out chat or diffusion follows from this TTS recipe.

Evidence & attribution

Source recipe Engine deployment revision

Imported reports retain their authorship and verification scope. Catalog checks do not establish inference quality or performance.